There's a potential issue here if ignored_layers contains an empty string. shard_prefix.startswith('') is always True, which would cause is_shard_skipped to be True for all layers, effectively disabling quantization for all fused layers. This could lead to silent failures and incorrect model behavior. It's safer to filter out empty strings from ignored_layers before checking startswith.

Similar to the change for fused layers, this logic is vulnerable to an empty string in ignored_layers. If ignored_layers contains '', prefix.startswith('') will always be True, causing all layers to be skipped from quantization. This can be prevented by ensuring ignored_layer is not an empty string before calling startswith.

Handle partial module prefixes anywhere in layer path

The new startswith checks still only match when the ignored layer string is at the very beginning of the module name. In the case described in the commit message (e.g. modules_to_not_convert contains "lm_head" but the actual prefix seen by quantization is "llm.lm_head"), prefix.startswith("lm_head") returns False, so the layer is quantized even though it was supposed to be skipped. As a result ms‑swift fp8 checkpoints with truncated prefixes will still attempt to quantize modules like llm.lm_head and fail to load correctly. The comparison needs to match ignored entries anywhere in the qualified name (e.g. with in/endswith) rather than only at the start.

Useful? React with 👍 / 👎.